Regarding Chinese AI laboratories employing distillation methods to gather insights from leading model developers, Y Combinator CEO Garry Tan expresses a desire for regulators to refrain from intervention. He suggests that U.S. AI labs might also engage in similar practices.

“I would take no action,” he told CNBC in an interview this week. “We could contend that an American distillation system should be established.”

He expanded on this to TechCrunch, indicating that he wishes for smaller, American open-weight AI labs to apply analogous training methods on U.S. frontier AI labs, thus providing the U.S. with a stronger array of open-weight options distinct from Chinese offerings.

Distillation involves a model developer extensively prompting another model to comprehend its functions and reasoning. It is a common, legitimate practice among AI labs to assist in training new models.

This week, Anthropic released its second report accusing Chinese labs of performing “illicit distillation assaults,” concealing their identities to distill without authorization and utilizing fraud and stolen credentials. Anthropic CEO Dario Amodei had previously urged U.S. regulators to take action against distillation.

It is noteworthy that the leader of Silicon Valley’s esteemed and prolific startup accelerator holds a different view.

To clarify, Tan does not endorse American AI labs using stolen credentials for distillation. He advocates for them to have open access. Indeed, his position is twofold. He believes it is an overreach for AI labs to control what their clients can do with the information derived from their models.

Furthermore, he points out that proprietary AI labs did not seek permission when they gathered vast amounts of human knowledge for their training. They notoriously assimilated numerous copyrighted materials without consent from those intellectual property owners.

“Restricting what users and customers can perform with API calls to closed-weight models feels limiting, and there is a role for the government here to acknowledge that access to intelligence trained on broadly available public data should be regarded more as a public resource rather than something confined behind restrictive terms of service,” he told TechCrunch when questioned about why American labs should also have the freedom to distill.

Tan, who is such an enthusiastic AI user that he once referred to himself as experiencing cyber psychosis, advocates for a balance between open-weight AI labs and frontier labs.

“They are leading the frontier and pushing it onward. We want that to be sustainable and a lasting business model,” he told CNBC. “Open-weight models should offer individuals freedom and access.”

To him, the ultimate AI doomsday scenario is if all the substantial capabilities of frontier AI were to reside with a single dominating, proprietary entity. “The worst case, the doomsday scenario for AI is having just one company,” he stated. “It possesses the best access to funding. It employs the top AI researchers. It monopolizes it and suddenly there’s one company that is all-encompassing. And that would be detrimental.”